MiniMax H3 US Models Pipe (CRT)

A MiniMax H3 workflow is a wall of models: a diffusion checkpoint, its turbo twin, a 32B-class Qwen3-VL text/image encoder, a video VAE, an audio VAE. The MiniMax H3 US Models Pipe (CRT) node bundles all of that into a single models_pipe, and then does something clever - it reads which mode and turbo state your Unified Sampler is actually running, and only loads (and only even evaluates) the model variant that run needs.

That "lazy" part is the whole selling point. H3 is a 33B model; you do not want to keep the REF2VA checkpoint, the FL2VA checkpoint and both turbo variants resident when a given run touches one of them. With this node, a plain T2V non-turbo run loads the FL2VA base model and nothing else. Flick the sampler's mode to R2V and flip Turbo on, and it fetches the REF2VA turbo model instead. The heavy Qwen3-VL CLIP and the two VAEs ride along the whole time because every mode needs them.

How the lazy loading works

The node receives two hidden inputs from ComfyUI - the full prompt graph (DYNPROMPT) and its own unique id (UNIQUE_ID) - and scans the graph for the Unified Sampler(s) connected to this pipe. From each sampler it reads workflow_mode and turbo, computes which model family that implies, and requests exactly those sockets. In subgraph-expanded dynamic-prompt runs where exact id matching fails, it falls back to loading the union of every sampler's needs so the run still gets the right weights. The author also wired IS_CHANGED to bust the cache when the mode or turbo toggle flips, so you don't get stale results.

Inputs and output

Required - the pieces every mode needs:

• vae - the MiniMax H3 video VAE.
• audio_vae - the audio VAE; it decodes generated audio and, in REF2VA, encodes your reference soundtracks.
• clip - the Qwen3-VL MiniMax text/image encoder.

Optional - the four lazy model sockets:

• fl2va_model - the FL2VA diffusion model (used by T2V and I2V).
• fl2va_turbo_model - the FL2VA base merged with the FL2VA Turbo LoRA.
• ref2va_model - the REF2VA diffusion model.
• ref2va_turbo_model - the REF2VA base merged with its Turbo LoRA.

Wire these from the pack's AutoDL nodes - MiniMax H3 Model (CRT AutoDL), MiniMax H3 CLIP (CRT AutoDL), the video and audio VAE loaders, and the two Turbo LoRA nodes - or any of your own loaders. Output is a single models_pipe into the Unified Sampler.

Install

Part of CRT-Nodes by PGCRT. ComfyUI Manager → search CRT-Nodes → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

Restart ComfyUI. Red/NaN nodes after an update: right-click → Fix node (recreate).

Gotchas

This node is deliberately loud about failure. If it can't find a Unified Sampler consuming this pipe, or a required socket is empty, it raises a clear error and prints a [CRT MiniMaxH3][DIAG] line to the console - the author chose that over silently caching a broken pipe, which would starve downstream samplers forever. So if you hit "could not find a Unified Sampler," check that the pipe output actually feeds a MiniMax H3 Unified Sampler (CRT) and that you've connected the model socket that mode needs. And remember: the first run downloads whatever's missing (big files, progress in the console), and the H3 license means this is only for you if you're outside the US, EU, UK and South Korea.